BACK PROPAGATION TECHNIQUE-MODELING CORTICAL COMPUTATION

点击翻译按钮获取中文摘要

It has recently been found that model neural networks trained with the back propagation procedure to do brain-like computations often develop hidden units with properties very similar to cortical neurons. A particularly clear example is our recent model of area 7a of monkey posterior parietal cortex. The response properties of the model hidden units match closely those of a class of neurons making up about half the units in area 7a. Empirical observations of this sort suggest that the back propagation paradigm might serve as a general technique for analyzing the mechanism of cortical computation. If this is true, it will be possible to make model networks with hidden units corresponding to neurons in many different cortical areas. The research proposed here is designed to explore this conjecture by extending our modeling efforts to a variety of other cortical areas and computation. In particular we will enhance the original area 7a model so it deals with three- dimensional representation, develop a primary visual cortex model using our previous observations on orientation and stereo, and build models of the sensory-motor integration processes thought to occur in the parietal lobe.